Mistral Large 4, Nicknamed Le Chonk, Is Now Available on Hugging Face
Mistral Large 4, Le Chonk, is a 1T open-weight multimodal model with 49B active, 1M context for cybersecurity, coding and fine-tuning, on Hugging Face.
Key Takeaways
- Mistral Large 4, nicknamed “Le Chonk,” is a multimodal, multilingual Mixture-of-Experts model with roughly 1T total parameters, 49B active parameters, and a 1M-token context window.
- It combines two sparsity mechanisms: MoE routing selects 4 of 128 experts per token, plus one shared expert, while sparse attention selects up to 2,048 key token positions per query; vision input is handled by a 1.6B Pixtral encoder.
- The model is especially positioned for cybersecurity, coding agents, and knowledge work, scoring 82% on the Artificial Analysis Cyber Index for vulnerability reproduction and patching, with strong coding-agent benchmark results.
- As an open-weight release under a Mistral Ecosystem Open-Weight License, it lets defenders run, inspect, and fine-tune the model locally, reducing dependence on API providers and enabling transparent safety evaluation.
- It ships with day-zero inference support for Transformers, vLLM, and llama.cpp, with FP8, NVFP4, and community-produced GGUF quantizations available for different deployment needs.
Mistral has released Mistral Large 4, codenamed "Le Chonk," a multimodal, multilingual Mixture-of-Experts model that takes text and images as input and generates text. The model carries roughly 1T total parameters with 49B active parameters and a 1M-token context window. It ships in two official variants—FP8 and NVFP4—with additional GGUF quantizations produced by Unsloth and the llama.cpp team. The weights are distributed under the Mistral Ecosystem Open-Weight License, an MIT derivative that places limitations on model-as-a-service offerings.
For readers tracking the open-weight ecosystem, the release matters less for its size than for what it enables: a frontier-class model that defenders can run, inspect, and fine-tune on their own infrastructure. Mistral positions Mistral Large 4 particularly for cybersecurity, coding agents, and multimodality.
Two Kinds of Sparsity in One Architecture
Mistral Large 4 pairs a large pool of experts with sparse attention and a Pixtral vision encoder. The language model selects a small group of experts for each token, while its attention indexer selects which earlier token positions to attend to. These are two distinct forms of sparsity: one selects feed-forward networks, the other selects context positions.
The text decoder has 72 layers, a hidden dimension of 6,144, and 64 attention heads. Its attention blocks combine Multi-head Latent Attention (MLA) with a DeepSeek Sparse Attention (DSA) indexer—a direction also seen in models such as GLM 5.3. MLA keeps a compressed KV representation, and DSA selects up to 2,048 key positions for each query. Each decoder block follows a 3:1 RoPE to NoPE pattern: three decoder layers employ RoPE and every fourth layer uses NoPE. The model also shares KV state across layers.
The first decoder layer uses a dense feed-forward network. The other 71 layers use MoE blocks, in which a softmax router selects four of the 128 experts for each token. One shared expert always runs alongside the selected experts, for a total of five active experts per token. For positional encoding, Mistral uses LeRope, a YaRN-scaled RoPE with one learned multiplier per frequency.
The vision side is handled by a 1.6B Pixtral encoder with 30 layers, a hidden dimension of 2,048, and 16 attention heads. It uses 14 × 14 patches and a patch-merge projector with spatial merge size 2.
Cybersecurity: Reproducing Vulnerabilities Is the Point
Mistral Large 4 scores 82% on the Artificial Analysis Cyber Index test, which asks a model to reproduce a real vulnerability in open-source software and then patch the software. Mistral reports that several closed frontier models score near zero on the same test because they refuse to attempt the work. That gap matters for defenders: fixing a vulnerability usually requires reproducing it first.
In Mistral's internal testing, the model was also useful for malware analysis, vulnerability prioritisation, and writing detection rules—tasks it was not explicitly trained for. The company mainly positions the model for defensive security work. Vulnerability research and incident response often look like attacks in model input, so closed-source filters can block them, and a team depending on an API can lose access in the middle of an incident. Open weights put both the usage policy and the deployment under the user's control.
The model still declines harmful requests. On cyber prompts from JailbreakBench, StrongREJECT, and AgentHarm, its average refusal rate is higher than any other open-weight model, according to Mistral. Open weights also let researchers measure refusal behaviour directly instead of relying on a provider's description. For community builders, the release opens downstream paths: fine-tuning smaller models with TRL on Mistral Large 4 traces that patch a specific code base, or using reproduce-then-patch tasks—which carry a verifiable reward—as RL environments to train smaller security models.
Benchmark Results Across Coding, Cyber, and Agents
As of October 29, 2026, Mistral Large 4 is the fifth best open-weight model on the Artificial Analysis Intelligence index. On coding, it achieves 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4. Its combined Coding Agent Index score of 49.8% places it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. The coding capabilities extend to knowledge work, with 59.9% on AutomationBench and 1,393 Elo on AA-Briefcase, which evaluate business workflows and long-horizon professional tasks.
On cybersecurity benchmarks, three results stand out. Mistral Large 4 ranks among the top five models globally on the Artificial Analysis Cyber Index, an independent evaluation of models' ability to find and fix software vulnerabilities. It achieves 82% on vulnerability reproduction and patching—the highest reported score on that test within the index—testing both investigation and remediation. And it solves 93% of a suite of 40 challenges drawn from security competitions on Cybench. The model also demonstrates 93.3% attack resistance on Lakera's B3 benchmark, adding evidence of robustness against prompt-injection attacks.
Inference Support from Day Zero
Mistral Large 4 shipped with day-0 support for transformers, vLLM, and llama.cpp. GGUF quants are available from GGUF and the llama.cpp team, and users can deploy remotely with Inference Endpoints or use serverless inference through Inference Providers.
In Transformers, the implementation provides AutoModelForMultimodalLM, with AutoProcessor preparing the conversation and image inputs. The model accepts low, medium, or high as reasoning effort. The default is high, and reasoning cannot be disabled—setting reasoning_effort to none would still use high.
vLLM supports the NVFP4 checkpoint, which has a smaller memory footprint and requires NVIDIA Blackwell GPUs. The model also ships with a DSpark drafter for additional speed-ups with vLLM. Llama.cpp and Llama App already support the model; after starting the server, users can head to localhost:9931 to chat through the built-in WebUI, add images for inference, and adjust the reasoning lever. The llama.cpp team ships IQ2_XXS, Q2_K, Q4_K_M, NVFP4, Q8_0, and BF16 quants for further conversions. The 2-bit quants run on a single Apple Silicon Ultra computer with 512 GB of unified RAM; the others require distributed inference and can run on multiple NVIDIA or AMD cards.
In practice, running Q2_K on a single M3 Ultra produced a prefill throughput of about 90 tokens per second and a generation speed of 15 tokens per second. On 4 x RTX6000 Pro, the observed results were: NVFP4 at 32K context size with prefill of 130–590 tokens per second and generation around 10 tokens per second; IQ2_XXS at 800K context size with prefill of 500 tokens per second and generation of 30 tokens per second at short contexts, dropping to 129 tokens per second prefill and 3 tokens per second generation at long contexts.
Fine-Tuning and Deployment Options
TRL support arrived on day zero. Upgrading to the latest versions of transformers, trl, and peft is all that is required. Mistral's example fine-tune teaches the model multilingual thinking, so the tuned model can reason in whatever language the system prompt requests; the script runs on the FP8 variant on a node of 8xB300s.
Mistral Large 4 is also available on Inference Endpoints with a ready-to-deploy recipe on 8 x NVIDIA H200 GPUs. The recipe runs vLLM with the FP8 weights split across all eight GPUs using tensor and expert parallelism, and uses the model's built-in DSpark draft head for faster generation through speculative decoding. It serves up to 131K tokens of context and processes up to 16 requests at a time. Observed results: a single request gets a time to first token of about 0.6 seconds for a 1K-token prompt and a generation speed of roughly 65 tokens per second, reaching 120–138 tokens per second on chat prompts where DSpark's drafts are accepted. Sixteen-thousand-token prompts, four at a time, are prefilled at about 4,300 tokens per second in total, with a time to first token of 8–9 seconds. Under load, the endpoint delivers about 480 tokens per second of total generation across 1,000 ShareGPT conversations. The endpoint exposes an OpenAI-compatible API; the reasoning_effort field controls how much the model reasons before answering, and the answer comes back in choices[0].message.content with reasoning in choices[0].message.reasoning. Tool calls, structured outputs, image inputs, and streaming are supported through standard Chat Completions fields.
For Dell AI Servers (NVIDIA H200, B200, GB200, or GB300), the dell-ai CLI installs an optimized recipe for Mistral Large 4, for example on an NVIDIA B300 node. More detail is available in the Dell Enterprise Hub.
What to Watch
The weights are available in the Mistral Large 4 collection, with GGUFs for local runs and Inference Endpoints for hosted deployment. The immediate question for security teams is whether an open-weight model that will attempt vulnerability reproduction changes their incident-response playbook; the immediate question for builders is how far fine-tunes on Mistral Large 4 traces can push smaller, cheaper security models. Mistral is inviting the community to share fine-tunes, quants, and other artefacts on the Hub.

Comments (0)